RoGemma2-9b-Instruct: A Specialized Romanian LLM
OpenLLM-Ro/RoGemma2-9b-Instruct is a 9 billion parameter instruction-tuned model developed by OpenLLM-Ro, marking the first open-source initiative to create a large language model specifically for Romanian. It is fine-tuned from Google's gemma-2-9b-it and trained on a comprehensive suite of Romanian instruction datasets, including RoAlpaca, RoDolly, and RoUltraChat.
Key Capabilities & Features
- Romanian Language Specialization: Optimized for natural language understanding and generation in Romanian.
- Instruction Following: Designed for assistant-like chat applications and responding to instructions.
- Research Focus: Intended for research use in Romanian NLP tasks.
- Diverse Training Data: Benefits from fine-tuning on multiple Romanian instruction datasets, enhancing its conversational and task-specific abilities.
Performance Highlights
While the base gemma-2-9b-it model often shows strong performance, RoGemma2-9b-Instruct demonstrates competitive results on Romanian-specific benchmarks. For instance, on the LaRoSeDa few-shot binary classification, it achieves 84.23% Macro F1, and on XQuAD few-shot, it reaches 49.22% EM and 66.33% F1. It consistently answers in Romanian, achieving 100% on RoCulturaBench and 160/160 on MT-Bench for Romanian responses.
Intended Use Cases
This model is ideal for research in Romanian natural language processing, particularly for developing conversational agents, instruction-following systems, and other NLP applications requiring strong Romanian language capabilities. Its instruct-tuned nature makes it suitable for assistant-like chat interactions.